# Kite

> Train robots faster. Augment datasets, train across models, and evaluate policies in simulation — on cloud GPUs, on demand.

- [Kite Platform API: augment robot datasets and train policies from your code](https://docs.kiteml.com/index.md): One small REST API for running Kite from your own code — augment robot datasets and train policies on cloud GPUs today, evaluate in simulation soon.
- [Quickstart: augment a LeRobot dataset with one call](https://docs.kiteml.com/quickstart.md): Go from an API key to a downloaded, augmented LeRobot dataset in a few minutes using the Kite Platform API and a single curl request.
- [Platform API overview: augment datasets and train policies on demand](https://docs.kiteml.com/platform-api/overview.md): A small, predictable REST API for running the Kite platform from your own code — dataset augmentation, policy training, and async runs on cloud GPUs.
- [API authentication: bearer tokens, scopes, and 401s](https://docs.kiteml.com/platform-api/authentication.md): Create a Kite API key and authenticate every Platform API request as a Bearer token, including key format, scopes, and 401 handling for invalid keys.
- [Dataset augmentations: create, poll, and download runs](https://docs.kiteml.com/platform-api/augmentation.md): Generate augmented robotics datasets programmatically — point Kite at a dataset, describe the change, and get back a ready-to-train Parquet dataset.
- [Training runs: fine-tune a policy on managed GPUs](https://docs.kiteml.com/platform-api/training-runs.md): Train robot policies programmatically — point Kite at a LeRobot dataset, pick your policies and a GPU tier, and get back trained checkpoints.
- [Digital twins: rebuild a robot episode as a MuJoCo scene](https://docs.kiteml.com/platform-api/twins.md): Point Kite at one episode of a LeRobot dataset and get back an interactable MuJoCo scene of the room, with the objects the robot handles built to size, delivered as one verified download.
- [RL runs: train a robot policy in simulation from a task spec](https://docs.kiteml.com/platform-api/rl-runs.md): Describe the behavior, robot, and budget as JSON. Kite trains the policy with reinforcement learning on a GPU and returns it as ONNX, with the MuJoCo scene it trained in, clips, and a verdict from measured checks.
- [Platform API reference: endpoints, methods, and paths](https://docs.kiteml.com/platform-api/api-reference.md): Every Kite Platform API endpoint at a glance — augmentation, training, and account routes, with methods, paths, and links to the OpenAPI spec.
- [CLI and MCP server: run Kite from a terminal or an AI agent](https://docs.kiteml.com/tools/cli-and-mcp.md): Install the kite CLI to augment datasets, build twins, and train RL policies from your terminal, or connect the Kite MCP server so Claude, Cursor, or any MCP client can do it for you.
- [Get Current Key](https://docs.kiteml.com/api-reference/api-keys/get-current-key.md): Introspect the caller's credentials: which key, which scopes, which user.
- [Estimate](https://docs.kiteml.com/api-reference/augmentations/estimate.md): Token cost estimate, billed per second of video per camera. Video augmentation bills up to 30 s of each episode; relighting bills every episode at full length, on the cameras it changes. Actual runs bill the real clip durations, so this is the price of the stated assumptions.
- [List Augmentations](https://docs.kiteml.com/api-reference/augmentations/list-augmentations.md)
- [Create Augmentation](https://docs.kiteml.com/api-reference/augmentations/create-augmentation.md)
- [Get Augmentation](https://docs.kiteml.com/api-reference/augmentations/get-augmentation.md)
- [Cancel Augmentation](https://docs.kiteml.com/api-reference/augmentations/cancel-augmentation.md)
- [List Episodes](https://docs.kiteml.com/api-reference/augmentations/list-episodes.md)
- [Get Episode](https://docs.kiteml.com/api-reference/augmentations/get-episode.md)
- [Get Episode Video](https://docs.kiteml.com/api-reference/augmentations/get-episode-video.md): One generated camera clip (mp4): 302 to a signed URL, or stream it through the API when signing is unavailable.
- [Download Output File](https://docs.kiteml.com/api-reference/augmentations/download-output-file.md): Download one file of a download-output dataset, streamed through the API. The fallback for `output.files[].url` when a direct download link isn't available; also a stable download route that always takes your API key.
- [Estimate](https://docs.kiteml.com/api-reference/training-runs/estimate.md): What training runs will be charged, in tokens, before you create them.
- [List Training Runs](https://docs.kiteml.com/api-reference/training-runs/list-training-runs.md): List your API-created training runs, newest first.
- [Create Training Runs](https://docs.kiteml.com/api-reference/training-runs/create-training-runs.md): Launch one run per policy. Returns a list of resources, one per policy.
- [Get Training Run](https://docs.kiteml.com/api-reference/training-runs/get-training-run.md): Status, live progress, metrics, and output for one run.
- [Get Training Run Logs](https://docs.kiteml.com/api-reference/training-runs/get-training-run-logs.md): The tail of a run's training log: what the trainer printed, oldest first.
- [List Checkpoints](https://docs.kiteml.com/api-reference/training-runs/list-checkpoints.md): Every checkpoint this run has saved, oldest first.
- [Download Checkpoint](https://docs.kiteml.com/api-reference/training-runs/download-checkpoint.md): Download one checkpoint as a ZIP of its `pretrained_model/` directory.
- [Cancel Training Run](https://docs.kiteml.com/api-reference/training-runs/cancel-training-run.md): Cancel a run and its GPU job. Terminal runs are a 409.
- [List Training Policies](https://docs.kiteml.com/api-reference/catalog/list-training-policies.md): Policy types you can train, and the smallest GPU tier each one needs.
- [List Hardware Tiers](https://docs.kiteml.com/api-reference/catalog/list-hardware-tiers.md): GPU tiers you can train on, with the token rate each one bills at.
- [Inspect Dataset](https://docs.kiteml.com/api-reference/catalog/inspect-dataset.md): Read a LeRobot dataset's metadata without starting a run.
- [Validate Twin](https://docs.kiteml.com/api-reference/twins/validate-twin.md): Everything a create checks, for free, without creating anything: the dataset resolves and has video, which camera will seed the room, and how long a twin typically takes.
- [List Twins](https://docs.kiteml.com/api-reference/twins/list-twins.md): Your twins, newest first, with the last status seen (GET one twin for live progress and its download).
- [Create Twin](https://docs.kiteml.com/api-reference/twins/create-twin.md): Build a twin from one episode of a Hugging Face LeRobot dataset. Only `source` is required.
- [Get Twin](https://docs.kiteml.com/api-reference/twins/get-twin.md): Live status, progress, each finished stage's measurement, and, once `succeeded`, `download`: a signed link to the whole scene as one .tar.gz (`curl -fL "$url" -o twin.tar.gz && tar xzf twin.tar.gz` gives `<twin id>/scene.xml` and `assets/`). Its file list is GET /v1/twins/{id}/files.
- [List Twin Files](https://docs.kiteml.com/api-reference/twins/list-twin-files.md): Every file in a succeeded twin's archive, by path, with its size and sha256. Paginated: pass the last `path` as `after`. A twin that has not succeeded is a 409 `twin_not_ready`.
- [Get Twin Archive](https://docs.kiteml.com/api-reference/twins/get-twin-archive.md): The finished scene as one .tar.gz: a 302 to a signed URL straight from storage, or streamed through this API when signing is unavailable. `curl -fL -H "Authorization: Bearer $KITE_API_KEY" .../archive -o twin.tar.gz`
- [Cancel Twin](https://docs.kiteml.com/api-reference/twins/cancel-twin.md): Stop a queued or processing twin. It is `canceling` until the build has actually stopped (seconds), then `canceled`, and twin.canceled is sent. The stages it finished stay cached for a later resume.
- [Resume Twin](https://docs.kiteml.com/api-reference/twins/resume-twin.md): Re-queue a failed or canceled twin. Every stage it finished is cached, so only what did not finish runs again.
- [Get Catalog](https://docs.kiteml.com/api-reference/rl-runs/get-catalog.md): Robots (with their verified command bands and measured throughput), objectives with every parameter's default and bounds, hardware tiers with their per-minute price, and the dated engines.
- [Plan Run](https://docs.kiteml.com/api-reference/rl-runs/plan-run.md): A task spec from a plain-English description, with the reasoning behind it. Changes nothing: pass the returned `task` and `budget` to POST /v1/rl_runs (after POST /v1/rl_runs/validate, ideally).
- [Validate Run](https://docs.kiteml.com/api-reference/rl-runs/validate-run.md): Everything a create checks, plus a dry run, for free and in seconds: the spec is compiled and three canned policies are rolled through it on a CPU — zero action, random actions, and, for a robot that has one, a public walk that passes the report's checks — reporting what each reward term pays each o…
- [Estimate Run](https://docs.kiteml.com/api-reference/rl-runs/estimate-run.md): Minutes and tokens for the create body, from throughput measured on this robot's own runs. Tokens are reserved at create and whatever training does not use is refunded.
- [List Runs](https://docs.kiteml.com/api-reference/rl-runs/list-runs.md): Your RL runs, newest first, with the last status seen (GET one run for live progress).
- [Create Run](https://docs.kiteml.com/api-reference/rl-runs/create-run.md): Train a policy. Only `robot` and `task.objective` are required; everything else has a verified default.
- [Get Run](https://docs.kiteml.com/api-reference/rl-runs/get-run.md): Live status and metrics; once packaged, the verdict and where the outputs are (`output`).
- [Get Metrics](https://docs.kiteml.com/api-reference/rl-runs/get-metrics.md): The training curve, every 5 iterations: mean reward per step, the same by reward term, episode length and throughput.
- [List Checkpoints](https://docs.kiteml.com/api-reference/rl-runs/list-checkpoints.md): Iterations the run saved (about every tenth of its budget, and the last). Any of them can seed a fork or a continue (`from.checkpoint`).
- [Download Checkpoint](https://docs.kiteml.com/api-reference/rl-runs/download-checkpoint.md): One saved iteration: `onnx` (the default: the deployable policy, raw observation in, action out) or `pt` (the rsl_rl checkpoint).
- [Get Report](https://docs.kiteml.com/api-reference/rl-runs/get-report.md): The verdict and how it was reached: each check with its value and limit, the metrics (24 worlds × 12 s, training noise on), the judge's read of the clip, and which checkpoint was chosen over which.
- [List Files](https://docs.kiteml.com/api-reference/rl-runs/list-files.md): Every file in the bundle, by path, with its size and sha256. Paginated: pass the last `path` as `after`.
- [Get File](https://docs.kiteml.com/api-reference/rl-runs/get-file.md): One file of the bundle, e.g. `policy/policy.onnx`, `media/video.mp4` or `mujoco/scene.xml`.
- [Get Archive](https://docs.kiteml.com/api-reference/rl-runs/get-archive.md): The whole bundle as one zip, `kiteml_<run id>.zip`, unpacking to `kiteml_<run id>/`. `curl -fLOJ -H "Authorization: Bearer $KITE_API_KEY" .../archive`
- [Cancel Run](https://docs.kiteml.com/api-reference/rl-runs/cancel-run.md): Stop a run. One that has started training keeps the policy it has so far: it shows `canceling` while that policy is packaged like a finished run's, then ends `canceled` with its outputs, charged for the minutes it trained. One that never started, or left nothing to package, is refunded in full.
- [Get Operation](https://docs.kiteml.com/api-reference/operations/get-operation.md)
- [List Webhook Endpoints](https://docs.kiteml.com/api-reference/webhooks/list-webhook-endpoints.md)
- [Create Webhook Endpoint](https://docs.kiteml.com/api-reference/webhooks/create-webhook-endpoint.md)
- [Delete Webhook Endpoint](https://docs.kiteml.com/api-reference/webhooks/delete-webhook-endpoint.md)
- [List Events](https://docs.kiteml.com/api-reference/webhooks/list-events.md): The durable event log (newest first) — the polling fallback for webhooks.
- [Get Usage](https://docs.kiteml.com/api-reference/usage/get-usage.md): Per-day, per-resource token usage (default: the last 30 days).
- [Get Credits](https://docs.kiteml.com/api-reference/usage/get-credits.md): Current token balance.

## OpenAPI Specs

- [openapi](/api-reference/openapi.json)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.